mtmd : Fix MinicpmV model converter and clip to avoid using hardcode. (#14750)

* Fix MinicpmV model converter and clip to avoid using hardcode.

* Code update for pr/14750

* Remove unused field, update script path in docs.

* Add version 5 for fallback code.

---------

Co-authored-by: lzhang <zhanglei@modelbest.cn>

tools/mtmd/clip-impl.h

@@ -44,6 +44,7 @@
 #define KEY_WIN_ATTN_PATTERN      "clip.vision.n_wa_pattern"
 #define KEY_ATTN_WINDOW_SIZE      "clip.vision.window_size"
 #define KEY_MINICPMV_VERSION      "clip.minicpmv_version"
+#define KEY_MINICPMV_QUERY_NUM    "clip.minicpmv_query_num"
 
 // audio-specific
 #define KEY_A_NUM_MEL_BINS      "clip.audio.num_mel_bins"

tools/mtmd/clip.cpp

@@ -201,6 +201,7 @@ struct clip_hparams {
     // legacy
     bool has_llava_projector = false;
     int minicpmv_version = 0;
+    int32_t minicpmv_query_num = 0;         // MiniCPM-V query number
 };
 
 struct clip_layer {
@@ -866,21 +867,8 @@ struct clip_graph {
             int n_embd = clip_n_mmproj_embd(ctx);
             const int d_head = 128;
             int n_head = n_embd/d_head;
-            int num_query = 96;
-            if (ctx->model.hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                num_query = 96;
-            } else if (ctx->model.hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                num_query = 64;
-            } else if (ctx->model.hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                num_query = 64;
-            }
-
+            // Use actual config value if available, otherwise fall back to hardcoded values
+            int num_query = ctx->model.hparams.minicpmv_query_num;
             ggml_tensor * Q = ggml_add(ctx0,
                 ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q),
                 model.mm_model_attn_q_b);
@@ -2138,7 +2126,19 @@ struct clip_model_loader {
                 get_u32(KEY_PATCH_SIZE, hparams.patch_size);
                 get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
                 get_i32(KEY_MINICPMV_VERSION, hparams.minicpmv_version, false); // legacy
-
+                get_u32(KEY_MINICPMV_QUERY_NUM, hparams.minicpmv_query_num, false);
+                if (hparams.minicpmv_query_num == 0) {
+                    // Fallback to hardcoded values for legacy models
+                    if (hparams.minicpmv_version == 3) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 4) {
+                        hparams.minicpmv_query_num = 64;
+                    } else if (hparams.minicpmv_version == 5) {
+                        hparams.minicpmv_query_num = 64;
+                    } else {
+                        hparams.minicpmv_query_num = 96;
+                    }
+                }
             } else if (is_audio) {
                 get_u32(KEY_A_NUM_MEL_BINS, hparams.n_mel_bins);
 
@@ -3556,20 +3556,23 @@ int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * im
             } break;
         case PROJECTOR_TYPE_MINICPMV:
             {
-                if (params.minicpmv_version == 2) {
-                    // MiniCPM-V 2.5
-                    n_patches_sq = 96;
-                } else if (params.minicpmv_version == 3) {
-                    // MiniCPM-V 2.6
-                    n_patches_sq = 64;
-                } else if (params.minicpmv_version == 4) {
-                    // MiniCPM-o 2.6
-                    n_patches_sq = 64;
-                } else if (params.minicpmv_version == 5) {
-                    // MiniCPM-V 4.0
-                    n_patches_sq = 64;
+                // Use actual config value if available, otherwise fall back to hardcoded values
+                if (params.minicpmv_query_num > 0) {
+                    n_patches_sq = params.minicpmv_query_num;
                 } else {
-                    GGML_ABORT("Unknown minicpmv version");
+                    // Fallback to hardcoded values for legacy models
+                    if (params.minicpmv_version == 2) {
+                        n_patches_sq = 96;
+                    } else if (params.minicpmv_version == 3) {
+                        n_patches_sq = 64;
+                    } else if (params.minicpmv_version == 4) {
+                        n_patches_sq = 64;
+                    } else if (params.minicpmv_version == 5) {
+                        // MiniCPM-V 4.0
+                        n_patches_sq = 64;
+                    } else {
+                        GGML_ABORT("Unknown minicpmv version");
+                    }
                 }
             } break;
         case PROJECTOR_TYPE_QWEN2VL:
@@ -4102,7 +4105,6 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
 }
 
 int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
-    const auto & hparams = ctx->model.hparams;
     switch (ctx->model.proj_type) {
         case PROJECTOR_TYPE_LDP:
             return ctx->model.mm_model_block_1_block_2_1_b->ne[0];
@@ -4114,20 +4116,7 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
         case PROJECTOR_TYPE_MLP_NORM:
             return ctx->model.mm_3_b->ne[0];
         case PROJECTOR_TYPE_MINICPMV:
-            if (hparams.minicpmv_version == 2) {
-                // MiniCPM-V 2.5
-                return 4096;
-            } else if (hparams.minicpmv_version == 3) {
-                // MiniCPM-V 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 4) {
-                // MiniCPM-o 2.6
-                return 3584;
-            } else if (hparams.minicpmv_version == 5) {
-                // MiniCPM-V 4.0
-                return 2560;
-            }
-            GGML_ABORT("Unknown minicpmv version");
+            return ctx->model.mm_model_proj->ne[0];
         case PROJECTOR_TYPE_GLM_EDGE:
             return ctx->model.mm_model_mlp_3_w->ne[1];
         case PROJECTOR_TYPE_QWEN2VL:

tools/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py

@@ -517,6 +517,16 @@ if args.use_f32:
 # output in the same directory as the model if output_dir is None
 dir_model = args.model_dir
 
+# Read config.json to get actual model configuration
+config_path = os.path.join(dir_model, "config.json")
+model_config = {}
+if os.path.isfile(config_path):
+    with open(config_path, "r", encoding="utf-8") as f:
+        model_config = json.load(f)
+    print(f"Loaded config from {config_path}")
+else:
+    print(f"Warning: config.json not found at {config_path}")
+
 # If minicpmv_projector is not specified but the default path exists, use the default path
 if args.minicpmv_projector is None:
     default_projector_path = os.path.join(dir_model, "minicpmv.projector")
@@ -555,37 +565,62 @@ if args.use_f32:
 #     processor = CLIPProcessor.from_pretrained(dir_model)
 
 minicpmv_version = args.minicpmv_version
-emb_dim = 4096
-block_count = 26
-if minicpmv_version == 1:  # MiniCPM-V 2.0
-    emb_dim = 2304
-    block_count = 26
-elif minicpmv_version == 2:  # MiniCPM-V 2.5
-    emb_dim = 4096
-    block_count = 27
-elif minicpmv_version == 3:  # MiniCPM-V 2.6
-    emb_dim = 3584
-    block_count = 27
-elif minicpmv_version == 4:  # MiniCPM-o 2.6
-    emb_dim = 3584
-    block_count = 27
-elif minicpmv_version == 5:  # MiniCPM-V 4.0
-    emb_dim = 2560
-    block_count = 27
 
-default_vision_config = {
-        "hidden_size": 1152,
-        "image_size": 980,
-        "intermediate_size": 4304,
-        "model_type": "idefics2",
-        "num_attention_heads": 16,
-        "num_hidden_layers": 27,
-        "patch_size": 14,
+# Use actual config values instead of hardcoded ones
+if model_config:
+    # For the projector/resampler, use the main model's hidden_size
+    emb_dim = model_config.get("hidden_size", 1536)
+
+    # For the vision model, use vision_config values
+    vision_config_dict = model_config.get("vision_config", {})
+    default_vision_config = {
+        "hidden_size": vision_config_dict.get("hidden_size", 1152),
+        "image_size": vision_config_dict.get("image_size", 980),
+        "intermediate_size": vision_config_dict.get("intermediate_size", 4304),
+        "model_type": vision_config_dict.get("model_type", "siglip"),
+        "num_attention_heads": vision_config_dict.get("num_attention_heads", 16),
+        "num_hidden_layers": vision_config_dict.get("num_hidden_layers", 27),
+        "patch_size": vision_config_dict.get("patch_size", 14),
     }
 
+    # Use vision model's num_hidden_layers for block_count
+    block_count = vision_config_dict.get("num_hidden_layers", 27)
+
+    print(f"Using config values: emb_dim={emb_dim}, block_count={block_count}")
+    print(f"Vision config: {default_vision_config}")
+else:
+    # Fallback to original hardcoded logic if config.json not found
+    emb_dim = 4096
+    block_count = 26
+    if minicpmv_version == 1:
+        emb_dim = 2304
+        block_count = 26
+    elif minicpmv_version == 2:
+        emb_dim = 4096
+        block_count = 27
+    elif minicpmv_version == 3:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 4:
+        emb_dim = 3584
+        block_count = 27
+    elif minicpmv_version == 5:
+        emb_dim = 2560
+        block_count = 27
+
+    default_vision_config = {
+            "hidden_size": 1152,
+            "image_size": 980,
+            "intermediate_size": 4304,
+            "model_type": "idefics2",
+            "num_attention_heads": 16,
+            "num_hidden_layers": 27,
+            "patch_size": 14,
+        }
+
 vision_config = Idefics2VisionConfig(**default_vision_config)
 model = Idefics2VisionTransformer(vision_config)
-if minicpmv_version == 3:
+if minicpmv_version == 3 or (model_config and model_config.get("vision_config", {}).get("model_type") == "siglip"):
     vision_config = SiglipVisionConfig(**default_vision_config)
     model = SiglipVisionTransformer(vision_config)
 elif minicpmv_version == 4:
@@ -644,16 +679,27 @@ else:
     fout.add_description("two-tower CLIP model")
 
 if has_vision_encoder:
-    # vision_model hparams
-    fout.add_uint32("clip.vision.image_size", 448)
-    fout.add_uint32("clip.vision.patch_size", 14)
-    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
-    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
+    # vision_model hparams - use actual config values
+    vision_image_size = model_config.get("image_size", 448) if model_config else 448
+    vision_patch_size = default_vision_config.get("patch_size", 14)
+    vision_hidden_size = default_vision_config.get("hidden_size", 1152)
+    vision_intermediate_size = default_vision_config.get("intermediate_size", 4304)
+    vision_attention_heads = default_vision_config.get("num_attention_heads", 16)
+
+    fout.add_uint32("clip.vision.image_size", vision_image_size)
+    fout.add_uint32("clip.vision.patch_size", vision_patch_size)
+    fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), vision_hidden_size)
+    fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), vision_intermediate_size)
     fout.add_uint32("clip.vision.projection_dim", 0)
-    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
+    fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), vision_attention_heads)
     fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
     fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
 
+    # Add MiniCPM-V specific parameters
+    query_num = model_config.get("query_num", 0) if model_config else 0
+    resampler_emb_dim = model_config.get("hidden_size", 0) if model_config else 0
+    fout.add_uint32("clip.minicpmv_query_num", query_num)
+
     if processor is not None:
         image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
         image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std

tools/mtmd/legacy-models/minicpmv-surgery.py

@@ -16,6 +16,8 @@ mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
 
 # store these tensors in a new dictionary and torch.save them
 projector = {name: checkpoint[name].float() for name in mm_tensors}
+if 'resampler.proj' in projector.keys() and hasattr(model.llm.config,'scale_emb') is True:
+    projector['resampler.proj'] = projector['resampler.proj'] / model.llm.config.scale_emb
 torch.save(projector, f"{args.model}/minicpmv.projector")
 
 clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]